128

10th International Conference of the International Institute for Infrastructure Resilience and Reconstruction (I3R2) 20–22 May 2014 Purdue University, West Lafayette, Indiana, USA Analysis and Classification of Volcanic Eruptions

Jithamala Caldera and S. C. Wirasinghe Department of Civil Engineering, University of Calgary

ABSTRACT Among natural disasters, volcanic eruptions are some of the most dangerous. The severity level of the most extreme volcanic eruption for which data is available can be categorized as Catastrophe Type II according to the scale introduced by Wirasinghe, Caldera, Durage, and Ruwanpura, (2013). However, an unusually large eruption of a “super volcano” can even cause a partial or full extinction. Aftermaths of a major eruption, such as climate effects, tsunami, and famine, severely impacts populations. Potential severity levels of volcanic eruptions are studied. A multidimensional scale for volcanic eruptions is investigated. Intensity, fatalities, affected population, impacted region, cost of damage, and GDP per capita, are some factors that can be considered to determine the severity level. Relationships among the factors are also considered. An analysis is conducted to identify the specific factors to be considered in the multidimensional scale. The extreme values of known historical eruptions of each volcano are studied in terms of fatalities. However, the study does not consider any secondary effect caused by the volcanic eruptions. The extreme values of fatalities from eruptions of 136 volcanoes are shown to be distributed as a 3 parameter Weibull (α = 0.33925, µ = 1, σ = 109.04) distribution.

1. INTRODUCTION the volcano is located, the effects will be felt globally or at least by a whole hemisphere”. The “fire from the earth,” volcano, is a crater of the earth’s crust. Hot magma, hot vapour, and gasses Volcanoes are found both on land and in the ocean escape through a vent when they erupt. Volcanoes (i.e., seamount volcanoes). About 94% of known have played a major role in the formation of the historical eruptions in the planet’s surface are Earth's atmosphere, ocean, and continents concentrated in linear belts (total length of 32,000 throughout history. They are one of the natural km and width of 100 km) which cover less than 0.6% disasters classified as a geophysical event (Low & of the Earth's surface (Sigurdsson, Houghton, Wirtz, 2010). Rymer, Stix, & McNutt, 1999). In addition, in Encyclopaedia of Volcanoes (Sigurdsson et al., Volcanoes grow by adding layers and height with the 1999) claims that 80% of the world's population lives accumulation of lava or ash. They can be classified in a nation with at least one Holocene volcano according to the level of activity as follows: active (active since the end of ice age, i.e., approximately (presently erupting), dormant (not presently erupting 11,700 years) and that the resources for dealing with but could at future date), and extinct (no eruptions in volcanic hazards are not evenly distributed. recorded history). In addition, volcanoes can be Moreover, he noted that because of the generality of categorized according to their shapes and sizes, the word “active,” the exact figure of the world's such as: compound volcanoes (complex of cones), active volcanoes cannot be accurately identified. stratovolcanoes (composite alternating layers of lava However, an approximate number of 1,500 and ash), somma volcanoes (a new central cone historically active or Holocene volcanoes are outgrowing the original caldera), and caldera identified on the Earth’s surface. On average, 50–60 (volcanic collapse crater [Siebert, Simkin, & volcanoes are active each year (Natural Kimberly, 2011]). Special types of volcanoes are also Environment Research Council, 2005). Mauna Loa distinguished, such as super volcanoes or hot spots. in Hawaii is the world's largest active volcano, rising Historical evidence of super volcanoes does not 13,677 feet above sea level; its top being over currently exist; however, evidence for these massive 28,000 feet above the nearby depth of the ocean phenomena have been observed in the geological floor. record. Examples of some hotspots are Hawaii, Yellowstone National Park (United States), Iceland, For the most part, volcanoes are primary disasters; Samoa, and Bermuda. Self (2006) has explained the however, they can also be secondary disasters when possible aftermath of super volcanic eruptions: “It is triggered by earthquakes. Volcanoes can, in turn, more likely that the Earth will next experience a result in secondary disasters such as tsunamis (e.g., super-eruption than an impact from a large meteorite 1883, Krakatau in Indonesia), famines (e.g., 1815, greater than 1 km in diameter. Depending on where Tambora in Indonesia), climate anomalies (e.g., 1815, Indonesia's Tambora causing June snow falls 129 and crop failures in New England, US), volcano Table 1. Volcanic Explosivity Index (VEI) criteria (Newhall & Self, collapses (e.g., 1792, Unzen in Japan), roof 1982). Source: Table 8 from Siebert, et.al., 2011 collapses, disease (e.g., 1991, Pinatubo in Philippines), and ash clouds (e.g., threat to air traffic, such as great circle routes to Japan over Alaska). Also, volcano eruptions can lead to pyroclastic flows (mixtures of hot gas and ash flowing at very high speeds [e.g., 100,200 km/h] leading to extreme heat and oxygen loss), lava flows (which are slow moving but can destroy houses, roads, and other structures), pollution (emission of strong poisonous gasses such as sulfur dioxide, hydrogen chloride, and hydrogen fluoride), mudflows (e.g., 1985, Ruiz, Colombia), ash flows (e.g., 1902, Mount Pelee, Martinique), and ash falls (causes respiratory problems and coverage of houses, buildings, roads, and crops with ash, [e.g., 1991, Chile's Cerro Hudson, Argentina, and 79 A.D., to the understanding of the classification of past Vesuvius in Italy). volcanic events by using the VEI scale on existing historical records. However, it was noted that the VEI The risk level of a volcano can be estimated by close was inadequate with respect to some aspects of monitoring. Nevertheless, the accurate estimation of disaster classification, for example, the climate volcano eruption is not currently possible, only the effect. approximate time of eruption can be estimated through regular monitoring. The estimated time to a Moreover, the same disaster can be interpreted by volcanic eruption may be given in hours, days, or several terms, according to the writer’s knowledge, can simply be a 30 second warning alarm experience, and personal feelings towards the (Wirasinghe, et.al, 2013). disaster. Several researchers have identified this as a major problem. Volcanology, unfortunately, has no Scientists measure the magnitude of eruptions by instrumentally determined magnitude scale, like that the Volcanic Explosivity Index (VEI). It is a general used by seismologists for earthquakes, and it is easy indicator of the explosive character of an eruption, as to understand why one observer’s “major” eruption shown in Table 1. Newhall and Self (1982) might be another’s “moderate” or even “small” event introduced this scale, and it distinguishes any (Siebert et.al., 2011). eruption in the range from 0 to 8. Intensity (column heights), magnitude (descriptive terms), and rate of Therefore, it is not uncommon for numerous records energy release during an eruption (blast durations) to exist for the same event, sometimes with differing are considered in this scale. When developing the conclusions. For example, the 1815 eruption of VEI scale, researchers identified that a quantitative Mount Tombora in Indonesia has different fatality or semiquantitative method for comparing eruptions records in different sources. “Victims from Volcanic was required. Eruptions: A Revised Database” (Tanguy, Ribière, Scarth, & Tjetjep, 1998) recorded the direct volcanic 2. PROBLEM STATEMENT effect fatalities as 11,000 (and other posteruption To predict severity levels of volcanic eruptions by famine and epidemic disease causing 49,000 utilizing available resources and technology, an fatalities). But it is given as 10,000 volcanic fatalities understanding of the mechanisms involved with (117,000 total fatalities) in the National Oceanic and volcanism is essential. The volcanism of a certain Atmospheric Administration (NOAA) database. region is characterized based on its history; Furthermore, the EM-DAT database records the therefore, it is necessary to document its full breadth. eruption of the Martinique Volcanic eruption on May Unless eruptions are documented at the time of their 8, 1902, as the most deadly volcanic event with occurrence, essential data required for prediction 30,000 fatalities. However, the NOAA database may be lost (Siebert, et al., 2011). At present, the records the same event with 28,000 fatalities. This number of reported eruptions is increasing as may be true for several reasons. Either some historic compared to the past; however, in general, records records may not be accurate or the disaster may be are incomplete. A few historical reports contain interpreted in different ways. For instance, one can some, but not all, of the necessary data; most count fatalities resulting directly from the volcano, or contain only a brief and often ambiguous description one can consider fatalities as a result of the of the eruptions (Newhall & Self, 1982). Volcanoes of aftermath as well (e.g., secondary disasters such as the World (Simkin, Siebert, McClelland, Bridge, starvation). Similarly, according to the EM-DAT Newhall, & Latter, 1981) has significantly contributed database, the Colombian eruption in 1985 resulted in the highest economic losses at around US $1 billion. 130

Whereas NOAA cites the highest economic losses With an increase in research and better early due to a volcanic eruption to be the 1980 Mount prediction technology, early warning systems and Saint Helen eruption in Washington, US, at US $2 hazard mitigation can also reduce the destructive billion. This is another example of inconsistency capacity of the eruption. among databases. Different scales have been introduced to distinguish Although historical inaccuracy of past records is the destructive capacity of a disaster. Disaster scope unavoidable, there should be a focus for avoiding (Gad-El-Hak, 2008) is an example of a disaster scale inaccuracies in the future. Hence, consistent for all types of disasters. Five levels (from Scope 1 to interpretation (primary/secondary disaster), proper 5) differentiate the severity of a disaster. Scopes are scale, good understanding of volcanoes, and an determined according to the number of casualties expanded recording system are required to and/or geographic area affected. Table 2 illustrates accomplish this goal. the disaster scope. The ranges proposed for casualties and the area affected appear to be The largest eruption might not be the most deadly arbitrary. eruption for several reasons. Other eruptions may have been as big as or bigger than the deadliest. For The fatality-based disaster scale (Wirasinghe et al., instance, current larger human populations than ever 2013) is another classification of natural disasters. before may increase the number of fatalities. On the Table 3 shows the fatality-based disaster scale which other hand, education levels of local residents about has categories using commonly used terminologies the disaster, existing technology, and available that describe various magnitudes. This scale has recourses may decrease the number of fatalities. been developed largely on the basis of fatality Table 2. Disaster scope statistics.

Casualties Area Affected The question is whether these scales can clearly Scope Disaster 2 (persons) (Km ) distinguish the severity levels of a disaster. Still, I Small <10 or <1 there is no right answer to this question because there are a lot of factors that need to be considered II Medium 10–100 or 1–10 when addressing the severity. However, lack of data III Large 100–1,000 or 10–100 prevents such sophisticated scale. If there is more IV Enormous 1,000–10,000 or 100–1,000 information available, a more advanced scale can be V Gargantuan >10,000 or >1,000 introduced. But to have good data, a better recording system and a more informative database is needed. Table 3. Fatality (F)-based disaster scale To have a better database, more precise terminology and a good scale are useful. Fatality Type Example Range In this paper, potential severity levels of volcanic eruptions are examined using available data. As Emergency 1 ≤ F < 10 A small landslide that kills one person discussed in the previous paragraph, the lack of data Disaster 10 ≤ F Edmonton tornado, Canada -1987 reduces the extent of the analysis. In total, 652 Type 1 < 100 that killed 27 people volcanic eruptions from 4360 B.C. to 2014 A.D. Disaster 100 ≤ F Thailand flood-2011 that resulted in a (February 1) obtained from the NOAA database for Type 2 < 1,000 total of 815 deaths both volcanic effects and total effects (volcano, Catastrophe 1,000 ≤ F Hurricane Katrina-2005, U.S.A that tsunami, earthquake, etc.) are considered. Five Type 1 < 10,000 killed 1833 people factors—number of fatalities, injuries, houses Catastrophe 10,000 ≤ F Tohuku earthquake and tsunami- damaged, missing people, and damage (in million Type 2 < 100,000 2011, Japan that killed 15882 people dollars)—have been studied. Data are available as Calamity 100,000 Haiti earthquake 2010 killed 316,000 numeric values and an interval variable (ordered Type 1 ≤ F < 1M people categories) for each factor. Although there are 652 Calamity 1M ≤ F China floods-1931 death toll recorded volcanic eruptions, all eruptions have no Type 2 < 10M >2,500,000 data value for at least two numerical variables. Cataclysm 10M ≤ F - Ordinal interval variables have more data values Type 1 < 100M compared to the corresponding individual numeric Super Volcano (e.g. Yellowstone) values; hence, the analysis is done using the interval Cataclysm 100M ≤ F Type 2 < 1B variable. Estimated deaths <1B 3. PARAMETERS THAT REFLECT THE Meteor strike (diameter > 1.5 Km) - SEVERITY OF A VOLCANIC EVENT Partial or 1B ≤ F estimated deaths :<1.5*109 Full < 10B This section is focused on identification of the Extinction Pandemic (Avian influenza) – estimated deaths : <2.8B parameters that reflect the severity of volcanic 131 eruptions. Intensity, fatalities, affected population, • Include interaction terms to the model (to impacted region, cost of damage, and GDP per address the multicollinearity effect) capita are some initial factors that could be o considered to determine the severity levels. Death*houses However, lack of historical data has limited the o Death*injuries analysis. Volcanic eruptions are measured using the VEI scale (0 to 8) which is an ordinal categorical o Houses*injuries variable. VEI is the best currently available factor • VEI grouping (lack of data in lower and that distinguishes one eruption from the other. Thus higher levels of VEI) the VEI scale is used to find the relationships between severity levels and other impact factors. o VEI (6,7,8->5) The SPSS software is used for the analysis using o VEI (0,1->1) (5,6,7,8->5) ordinal logistic regression. Link Function for logit model: All factors considered are ordinal variables; ProbabilityofVEI ≤ j therefore, spearman's rho correlation coefficient (ρ) Log =α+βx ; is used to observe the correlation. Each ordinal ProbabilityofVEI > j interval variable for “direct volcanic effects” shows an where j = 1, 2, 3, 4, 5. excellent linear relationship with the corresponding variable for “total effects” with ρ > 0.9 for all pairs. Ordinal interval variables of death, injuries, and Thus, direct volcanic effects alone can be used to houses damaged individually have formed a good explain the relationship with other factors. ordinal regression model with VEI. The best models are given when the link function is logit, as shown in Lack of data highly influenced the model selection the above equation, that is, with the assumption that even within the volcano effects. All except one the residuals are logistically distributed and VEI is eruption, from 652 that are recorded, are missing grouped (VEI 0,1 as VEI 1 and VEI 5,6,7,8 as VEI 5). one or more data for categorical variable. The VEI Therefore, the best selected models are death, scale is not defined for 90 eruptions. Correlation injuries, and houses damaged individually with amongst the volcanic impact variables has been grouped VEI scale. All the p-values corresponding to observed. Damage in millions of dollars has a very the estimated parameters are less than 0.1 in these good linear relationship with houses damaged (ρ = models. The estimated values represent the 0.9). One variable stayed in the model while the parameters α (threshold) and β (location) in the other is omitted because of the high correlation. above equation with a 90% confidence level. In Damage in millions of dollars has a close ordinal logistic regression, it is assumed that each relationship with time and inflation, and, thus, is hard level of VEI is parallel to the other. To select the best to estimate. Hence, it is omitted from the model. The models, the following hypothesis tests have been number of missing people and number of deaths are conducted with 95% confidence level: also highly correlated (ρ = 0.9). It can be observed that the number of pair wise data is fairly low with • Tests of parallel lines presence of the number of missing people. It may • explain the higher ρ value for some pairs. Therefore, Goodness of fit tests the number of missing people is also omitted from • Overall model fits the model. Other pairs, for example, deaths and houses damaged, are not highly correlated but have The models also improved when testing the data a moderate to good relationship (0.5 ≤ ρ < 0.75). only for the last 114 years, though the combination of VEI grouping and data for the last 114 years is not Different approaches have been tried to select a an improvement. The lack of data was felt again good relationship between VEI and the other when testing the model for the last 32 years. The variables. Some of the approaches are: accuracy of the assigned VEI scale for volcanic • Different link function (logit, probit, etc.) eruptions could have been tested through this approach if sufficient data existed. Moreover, one • Log transformation of death, house, injuries variable becomes significant with the presence of another variable because of multicollinearity • Different periods between two variables (e.g., injuries become o Last 32 years (after 1982), after the significant with the presence of deaths, houses VEI scale is introduced become significant with the presence of deaths and houses become significant with the presence of o Last 114 years, after 1900 injuries). The multicollinearity effect remains the o Last 514 years, after 1500 same for all applied approaches; hence, the 132 combination of several impact variables could not be Catastrophe Type 2. Table 5 shows the fatality-based achieved as expected. disaster scale for some volcanic eruptions. Volcanic eruptions can vary from Emergency to the Calamity The results highlight the fact that individual variables Type 1 level. However, an unusually large (super of death, injuries, and number of houses damaged are better than the combinations of the above variables in explaining the relationship with VEI. In other words, the use of “or” between the variables is better than the use of “and” when describing the relationship between severity of the volcanic eruptions and the above concerned variables. Only the number of fatalities has been discussed in the following section out of the number of houses damaged and number of people injured. 4. SEVIRITY LEVEL BOUNDARIES This section examined how the severity levels are spread with the number of fatalities, specifically the intervals or ranges or boundaries of fatalities which differentiate one severity level from the other. There are 236 volcanoes in the NOAA database which have at least one eruption. For extreme value Figure 1. Histogram of extreme fatalities for volcano effects and data analysis, only the maximum fatality recorded for the fitted Weibull (3P) density (dash line) different volcanoes is considered. For instance, in Table 4. Probability of an eruption to be of the given type the volcanic effects for the Mount Tombora 1815 eruption record 10,000, Mount Krakatoa 1883 Type Fatality Range Probability eruption record 2,000, etc. Likewise the maximum Emergency 1 ≤ F < 10 0.348852 fatality records for different volcanoes are selected. All other records for Mount Krakatoa except the 1883 Disaster Type 1 10 ≤ F < 100 0.271215 eruptions are not considered. All the records which Disaster Type 2 100 ≤ F < 1,000 0.259911 1,000 ≤ F < do not have at least one fatality are also not Catastrophe Type 1 0.110283 considered because the cells which are blank in the 10,000 10,000 ≤ F < database either represent no fatality or no record Catastrophe Type 2 0.009699 100,000 found. Then, extreme fatality recorded eruptions for 100,000 ≤ F < Calamity Type 1 0.000040 136 volcanoes are shown to be distributed as a 3 1M parameter Weibull (α = 0.33925, µ = 1, σ = 109.04) Calamity Type 2 1M ≤ F < 10M 0 distribution with a sample mean of 1,202.81, a Cataclysm Type 1 10M ≤ F < 100M 0 sample variance of 4,251.75, and a maximum of 30,000. Other extreme value distributions give Cataclysm Type 2 100M ≤ F < 1B 0 Partial or Full unrealistic values. Figure 1 shows the histogram of 1B ≤ F < 10B 0 the extreme fatality volcano effects and the fitted Extinction

Weibull (3P) density (dashed line). Table 5. Fatality-based disaster scale for volcano The considered dataset of volcanic effects followed the fatality-based disaster scale for all types of Example natural disasters (Wirasinghe et al., 2013). According Type to the fitted Weibull distribution for volcanic Year Volcano Country Fatalities eruptions, 35% of the eruptions are Emergency type; 27% and 26% of eruptions are Disaster Type 1 and Emergency 2011 Nabro Eritrea 7 2, respectively; and 11% and 1% of eruptions are Disaster 1975 Marapi Indonesia 80 Catastrophe Type 1 and 2, respectively. Wirasinghe Type 1 et. al. (2013) stated that volcanic eruptions go up to Disaster Catastrophe Type 2 level. However, four in 100,000 1991 Pinatubo Philippines 450 Type 2 eruptions have the ability to reach the Calamity Type Papua Catastrophe 1, according to the fitted Weibull (3P) distribution, 1951 Lamington New 2942 Type 1 though there is no evidence in recorded history. The Guinea severity level of the most extreme volcanic eruption Catastrophe 1985 Ruiz Colombia 23080 for which data is available (450 A.D., Ilopango, El Type 2 Salvador, 30,000 fatalities) can be categorized as 133 volcanic) eruption has the potential to exceed the Newhall, C. G., & Self, S. (1982). The volcanic above mentioned levels. They can cause a calamity explosivity index (VEI) an estimate of explosive or even a partial or full extinction. magnitude for historical volcanism. Journal of Geophysical Research: Oceans, 87(C2), 1231– 5. CONCLUSION 1238. http://dx.doi.org/10.1029/JC087iC02p01231 This paper presents the initial attempts to develop a Self, S. (2006). The effects and consequences of multidimensional scale for volcanic eruptions. The very large explosive volcanic eruptions. analysis is limited to five variables: number of Philosophical Transactions of the Royal Society A: fatalities, number of missing people, number of Mathematical, Physical and Engineering Sciences, injuries, number of houses damaged, and damage in 364(1845), 2073–2097. http://dx.doi.org/ million dollars. Extensions of this work are expected 10.1098/rsta.2006.1814 to be reported in future. Siebert, L., Simkin, T., & Kimberly, P. (2011). REFERENCES Volcanoes of the world (3rd ed.). Berkeley, CA: University of California Press. Centre for Research on the Epidemiology of Disasters. (n.d.). EM-DAT: The international Sigurdsson, H., Houghton, B., Rymer, H., Stix, J., & disaster database. Retrieved from McNutt, S. (1999). Encyclopaedia of volcanoes. http://www.emdat.be/database Burlington, MA: Academic Press. Gad-El-Hak, M. (2008). The art and science of large- Simkin, T., Siebert, L., McClelland, L., Bridge, D., scale disasters. In M. Gad-El-Hak (Ed.), Large- Newhall, C. G., & Latter, J. H. (1981). Volcanoes of scale disasters (pp. 5–68). New York, NY: the world: A regional directory, gazetteer, and Cambridge University Press. chronology of volcanism during the last 10,000 years. Stroudsburg, PA: Hutchinson Ross Low, P., & Wirtz, A. (June 2010). Structure and Publishing. needs of global loss databases about natural disaster. Paper presented at the IRDC Davos Tanguy, J. -C., Ribière, C., Scarth, A., & Tjetjep, W. 2010: Third International Disaster and Risk S. (1998). Victims from volcanic eruptions: A Conference, Davos, Switzerland. revised database. Bulletin of Volcanology, 60(2), 137–144. http://dx.doi.org/10.1007/s004450050222 National Geophysical Data Center. (n.d.). Volcano events search [Data file]. Available from Wirasinghe, S. C., Caldera, H. J., Durage S. W., & http://www.ngdc.noaa.gov/hazard/ Ruwanpura, J. Y. (2013). Preliminary analysis and classification of natural disasters. In Proceedings of Natural Environment Research Council. (2005). the ninth annual conference of the International Natural hazards scientific certainties and Institute for Infrastructure, Renewal and uncertainties. Swindon, UK: Natural Environment Reconstruction. Brisbane, Australia: Queensland Research Council. University of Technology.